Non-Experts Shape Modeling for Dummies

Transcription

1 Non-Experts Shape Modeling for Dummies (Modeling with Interchangeable Parts) Alla Sheffer (joint work with Vladislav Kraevoy & Dan Julius)

2 Motivation - Easy creation of 3D Content Currently 3D modeling requires a lot of time & expertise Observations: Practical modeling limited to small set of classes Models have intuitive breakdown into interchangeable parts Can create rich & detailed models by shuffling parts n models with m parts n m new models

3 Shuffler Modeling System

4 Modeling System

5 Shuffler Fast & Trivial to use Mouse click based No geometric input from user No user parameters Operates on models of generic topology multiple components, any genus, boundaries, non-manifold

6 Under the hood... I. Meaningful Segmentation II. Part Correspondence III. Shuffling: alignment & blending

7 Previous Work Modeling From scratch Traditional Tools [Maya, 3DMax] + Create rich & complex models time consuming, requires expertise Sketching [Igarashi 99] + Fast & easy Good for simple models

8 Previous Work Modeling Reuse + Create complex models Blending [Allen 03, Schreiner 04, Kraevoy 04] Users must specify feature correspondences Composition [Yu 04, Sorkine 04] Users must cut & position parts Modeling by Example [Funkhouser 04] + Find & position parts based on similarity Users must cut models

9 Previous Work - Segmentation Meaningful convex [Hoffman & Richards 84] Feature characteristics [Mortara 04, Attene 06] Exact convex [Chazelle 95] oversegments Parts from farthest points [Katz 03,Zhang 05,Katz 05] Distance to convex-hull [Lien 06] Typically require per model parameter fine-tuning

10 Previous work - Correspondences Use for matching score [Cornea 05, Sundar 03] Do not disambiguate symmetries Incomplete matches [Sundar 03] [Cornea 05] Topology restrictions (genus 0) Some assume rigid registration Partial matching [Funkhouser 04] single part [Novotni 06]

11 Segmentation I. Meaningful Segmentation II. Part Correspondence III. Shuffling: alignment & blending

12 Metrics Segmentation quality 1. Convexity 2. Compactness 3. Cost 4. (α the same for all models)

13 Algorithm 1. Patch generation hole filling 2. Seed generation 3. Patch growing & reseeding Use convexity bound 4. Repeat till no holes left

14 Segmentation

15 Hierarchical segmentation Bottom up Merge patches as long as combined patch satisfies a coarse threshold

16 Segmentation I. Meaningful Segmentation II. Part Correspondence III. Shuffling: alignment & blending

17 Correspondence Many-to-Many Number of groups unknown a priori Groups represent meaningful components Cost function Compare match quality across different groupings & different number of groups

18 Cost Function Group convexity measures grouping quality Measure similarity in terms of mid-point graph distances supports pose variation Mid-point graphs To resolve symmetries use Euclidean distances & volume

19 Part Matching - Search Use hierarchical segmentation Establish correspondences on coarse level Use stochastic local search iterate random initial guess selection local search

20 Part Matching - Search Do registration based on results Resolves global symmetries Needed for shuffling Refine correspondences using fine segmentation level Use group splits

21 Part Correspondence

22 Segmentation I. Meaningful Segmentation II. Part Correspondence III. Shuffling: alignment & blending

23 Shuffling Positioning global rotation & scale closest point alignment of matching group centers Position individual parts use common group graph

24 Shuffling Blending & Merging Extend meshes to form overlap region Perform zippering [Turk 94] & geometry blending

25 Shuffling - Results

26 Results - Example I. Meaningful Segmentation II. Part Correspondence III. Shuffling: alignment & blending

27 Shuffling - Results All the models created in less than a minute

28 Results Segmentation & Matching

29 Results Segmentation & Matching

30 Segmentation (CAD) - Comparison

31 Comparison (natural models)

32 Segmentation Comparison Comparison with [Lien 06]

33 Summary New modeling system for dummies Button click interface Robust segmentation algorithm Outperforms existing methods Method for matching interchangeable parts Computes complete correspondences use grouping Disambiguates symmetric parts No per-model parameters No topology restrictions

34 Future/Issues Segmentation Symmetry Straighter boundaries Matching Better understanding of correlation between human perception and geometric measures

35 Thank You!!! Questions?